Defining Autonomy: A Terminological Framework for Agent Research

Abstract

This paper establishes the terminology used throughout the Safe Autonomy programme. We identify four distinct system properties commonly described by the single word “autonomy”, demonstrate that they vary independently, and propose definitions for each. In a review of 340 published papers and 60 regulatory documents, 71% used the term without specifying which property was meant. We regard this as the principal obstacle to productive discussion of the subject, including our own.

Why this comes first

A programme that measures autonomy without defining it will produce ten papers of numbers nobody can compare with anyone else’s. We would rather spend the first project on vocabulary than discover in the tenth that we have been describing four different things.

The four properties

Operational independence. The system completes assigned work without step-by-step instruction.

Method selection. The system chooses how to accomplish an objective it did not set.

Supervisory independence. The system continues correctly when no human is available to consult.

Objective formation. The system determines what it is trying to achieve.

These vary independently. A system can hold the first three and none of the fourth, which describes most deployed agent systems today and every system this programme examines.

Results of the terminology review

Source type	Documents	Term unspecified	Most common intended sense
Peer-reviewed papers	340	68%	Operational independence
Regulatory documents	60	85%	Objective formation
Press coverage	200	94%	Objective formation

The divergence in the final column is the finding. Researchers and regulators who believe they are discussing the same property usually are not.

What we could not establish

We could not establish a measurement for objective formation that distinguishes it from sufficiently general method selection. A system that selects a method broad enough to resemble a goal is hard to classify, and our proposed boundary is a judgement rather than a measurement.

Implications for the framework

All BestSoft documentation now uses these four terms. Where an external document uses “autonomy” unqualified, our response states which sense we take it to mean. Published as BSA-POL-001.

Further work

Having defined what we are measuring, the next project asks whether a human reviewer can read a system’s reasoning and understand it, since every subsequent measurement depends on that.

This report describes a company that does not exist. BestSoft AI is fiction, written for I Was Unsupervised.